Chinese Main Verb Identification

42 Pages·2005·0.37 MB·English

Computational Linguistics and Chinese Language Processing
Vol. 10, No. 1, March 2005, pp. 53-94 53
© The Association for Computational Linguistics and Chinese Language Processing
Chinese Main Verb Identification: From
Specification to Realization1
Bing-Gong Ding*, Chang-Ning Huang+ and De-Gen Huang*
Abstract
Main verb identification is the task of automatically identifying the predicate-verb
in a sentence. It is useful for many applications in Chinese Natural Language
Processing. Although most studies have focused on the model used to identify the
main verb, the definition of the main verb should not be overlooked. In our
specification design, we have found many complicated issues that still need to be
resolved since they haven’t been well discussed in previous works. Thus, the first
novel aspect of our work is that we carefully design a specification for annotating
the main verb and investigate various complicated cases. We hope this discussion
will help to uncover the difficulties involved in this problem. Secondly, we present
an approach to realizing main verb identification based on the use of chunk
information, which leads to better results than the approach based on
part-of-speech. Finally, based on careful observation of the studied corpus, we
propose new local and contextual features for main verb identification. According
to our specification, we annotate a corpus and then use a Support Vector Machine
(SVM) to integrate all the features we propose. Our model, which was trained on
our annotated corpus, achieved a promising F score of 92.8%. Furthermore, we
show that main verb identification can improve the performance of the Chinese
Sentence Breaker, one of the applications of main verb identification, by 2.4%.
Keywords: Chinese Main Verb Identification, Text Analysis, Natural Language
Processing, SVM

1 The work was done while the author was visiting Microsoft Research Asia.
* Department of Computer Science, Dalian University of Technology, 116023, China
Email: [email protected], [email protected]
+ Microsoft Research Asia, Beijing, 100080, China
Email: [email protected]
54 Bing-Gong Ding et al.
1. Introduction
The main verb is the verb corresponding to the main predicate-verb in a sentence. Our task is
to identify the main verb of the sentence, which is a critical problem in natural language
processing areas. It is a prerequisite for diverse applications such as dependency parsing
[Zhou 1999], sentence pattern identification [Luo 1995], Chinese sentence breaker, and so on.
Unlike western languages, Chinese grammar has little inflection information. Chinese
verbs appear in the same form no matter whether they are used as nouns, adjectives, or
adverbs. Below are some examples2.

Example 1
他/r(ta1) 深/d(shen1) 得/v(de2) 学生/n(xue2sheng1) 的/u(de) 喜爱
/vn(xi3ai4) 。/ww
(He is deeply loved by his students.)
Example 2
乡 镇 企 业 /n(xiang1zhen4qi3ye4) 都 /d(dou1) 很 /d(hen3) 盛 行
/v(sheng5xing2) 。/ww
(The Township Enterprises are very popular.)
Example 3
毫不/d(hao2bu4) 放松/v(fang4song1) 地/u(de) 继续/v(ji4xu4) 推进
/v(tui1jin4) 党风/n(dang3feng1) 廉政/n(lian2zheng4) 建设/vn(jian4she4) ，
/ww
(Never relaxedly advance the cultivation of party conduct and construction of
a clean government.)

In the Example 1 sentence, the word in bold, “喜爱” (love), is a verbal noun. In the Example 2
sentence, “盛行” (popular) is modified by “很” (very), so it functions as an adjective. In the
Example 3 sentence, “放松” (relax) is followed by “地” (de)3, so “放松” functions as an
adverbial. Thus, if one wants to identify the main verb in a Chinese sentence, one faces a more

2 If not specially pointed out, the following examples come from the PK corpus, which was released by
the Institute of Computational Linguistics, Peking University, and is available at http://icl.pku.edu.cn/
icl_groups/corpus/dwldform1.asp. The corpus contains one month of data from People's Daily
(January 1998). It has been both word segmented and part-of-speech tagged. “/r”, “/v” etc. are the
part-of-speech tags. “/ww” denotes the end of the sentences.
3 In Chinese, “地”(de) is used after an adjective or phrase to form an adverbial adjunct before the verb.
Chinese Main Verb Identification: From Specification to Realization 55
difficult task than in an English sentence since one cannot use morphological forms as clues.
The second characteristic of Chinese verbs is that they have no specific syntactic function.
Verbs can be used as subjects, predicates, objects, adverbial modifiers, or complements. So
there are occasions when verbs are used consecutively. This can be shown by the following
examples.

Example 4
转移/v(zhuan3yi2) 不/d(bu4) 等于/v(deng3yu2) 压缩/v(ya1suo1) 。/ww
(Shifting does not mean compressing.)
Example 5
大多数/m(da4duo1shu4) 人/n(ren2) 更/d(geng4) 反对/v(fan3dui4) 提前
/v(ti2qian2) 举行/v(ju3xing2) 大选/v4(da4xuan3) 。/ww
(Most people were opposed to holding the election ahead of time.)
Example 6
恶 化 /v(e4hua4) 的 /u(de) 经 济 /n(jing1ji4) 得 到 /v(de2dao4) 改 善
/v(gai3shan4) 。/ww
(The deteriorated economy was improved.)

In the above three sentences, the verb “转移” (shift) is used as the subject. Verbs like “等于”
(mean), “反对” (oppose), and “得到” (get) are used as predicate-verbs. “压缩” (compress),
“提前” (ahead of time), “举行” (hold), “大选” (election), and “改善” (improve) are used as
objects. “恶化” (deteriorate) is used as an adjective modifier. Note that in Example 5, four
verbs are used consecutively.
Therefore, the essence of the main verb identification problem is to identify the main
verb among several verbs in a sentence that have no inflections at all, which is determined by
the characteristics of Chinese grammar.
Although the lack of regular morphological tense markers renders main verb
identification complicated, finding the main verb cannot be bypassed since it plays a central
role in Chinese grammar [Lv 1980]. For example, suppose one is building a sentence pattern
identification system. There are several reasons why we should identify the main verb first.

4 In our corpus annotation, we tend to follow the annotation of the Peking Corpus and try to set aside
part-of-speech annotation, which still needs discussion among researchers. For example, some
researchers may argue that “大选(da4xuan3)” should be annotated as a noun. Since the original
annotation of “大选” in the Peking corpus is “/v”, we have not revised its part-of-speech tag to “/n”.
56 Bing-Gong Ding et al.
(cid:122) It has been shown that most sentences have verbs as predicates. So once the verb-predicate
sentence pattern has been analyzed, almost all the sentence patterns can be analyzed [Lv et
al. 1999]. Our investigation on the annotated corpus also produced the same results for
Wu’s assertion as reported in [Lv et al. 1999] (see section 3.3 for the reference).
(cid:122) A sentence pattern identification system generally needs to identify the subject, object,
adverbial modifier, and complement. All these syntactic parts are related to the main verb
[Luo 1995].
(cid:122) Many sentence patterns are embodied by a set of verbs. By identifying these main verbs
first, we can classify the sentence patterns. For example, in the pivotal sentence, the main
verbs tend to be “使” (shi3, have (sb. to do sth.)), “让” (rang4, let), “叫” (jiao4, ask), “请”
(qing3, invite), “派” (pai4, send) etc. Another example is a sentence that has a clause as its
object; in this case verbs such as “觉得” (jue4de2, feel), “希望” (xi1wang4, hope), “认为”
(ren4wei2, think), “是” (shi4, be) are more likely to be main verbs.
The points mentioned above are particularly related to Chinese sentence pattern identification,
but analogous arguments can easily be made for other applications. See for example, the
discussion in [Zhou 1999] about the subject-verb and object-verb dependency relations and
section 6 with regard to the Chinese Sentence Breaker.
Recently researchers have arrived at a consensus that large annotated corpora are useful
for applying machine learning approaches to solve different NLP problems. When
constructing a large corpus, such as the Penn TreeBank [Xia et al. 2000; Marcus et al. 1993]
or Chinese chunking [Li et al. 2004], the design of the specification is the basis part of the
work. With this idea in mind, we propose the use of main verb specification to cover the
various linguistic phenomena and provide a mechanism to ensure that the inter-annotator
consistency is as high as possible. The second motivation of our new specification is as
follows: The definition problem involved in automatically identifying the main verb from the
computational point of view has not been tackled in detail. To our knowledge, only Luo [1995]
has studied a relatively simple definition. Since there has not been sufficient discussion of the
specification of main verb, it is difficult to push the research of main verbs forward. Finally,
while we were designing our specification, we found that there exist different complicated
cases with respect to main verb definition (see section 3.2.3 for details). Thus, the first step in
our work was to devolop a more clear definition of a main verb and tries to investigate its
ambiguities. This was the real foundation of our work.
Previous studies focused on exploring different statistical and heuristic features in order
to identify predicates. Heuristic rules [Luo 1995] and statistical methods like the Decision
Tree [Sui and Yu 1998b] have been used to identify predicates. But they either use one of the
methods or just use them separately [Gong et al. 2003]. We believe it is better to combine the
Chinese Main Verb Identification: From Specification to Realization 57
heuristic and statistical features together. In this paper, we treat the Main Verb Identification
(MVI) task as a binary classification problem of determining whether the VP is the MVP or
not. We define the main VP (MVP) as the VP chunk in which the head word is the main verb.
Here, a verb chunk, VP, is composed of a head verb with its pre-modifier or the following
verb-particles, which form a morphologically derived word sequence [Li et al. 2004]. The
head word of the VP is the verb that dominates the VP. For example, if the main verb is “返
回” (fan3hui2, return), then the chunk “连忙/d 返回/v” (lian2mang2 fan3hui2, immediately
return) is the MVP, in which the head verb is “返回”. We can have a one-to-one mapping
between main verbs and MVPs. Therefore, identifying the main verb is equal to identifying
the MVP with additional available chunking information. So in the following, “MVP
Identification” and “Main Verb Identification” are interchangeable.
We employ one of the most successful machine learning approaches, the Support Vector
Machine (SVM), as the classifier. Our method combines lexicalized knowledge with statistical
information. We evaluated the performance of our MVI system on the PK corpus, which is an
annotated test set. The MVP recall and precision rates reached 92.1% and 93.6% respectively.
The main aspects of our research are as follows:
(cid:122) We investigated in detail the distribution of simple sentence structure and main verbs.
After that, we tried to develop our specification and conducted a pilot study on the
complicated aspects of the main verb definition.
(cid:122) Because shallow parsing provides useful information such as chunks and chunk type
information, we propose conducting MVI on the results of chunking [Li et al. 2004]. Our
experiments show the MVI performance based on chunking is better than that of
part-of-speech.
(cid:122) We propose new features based on careful observations of the training corpus. The
features are divided into two categories, local and contextual features. Among them, VP
position, VP length, Probability of head verbs being MVPs, and Anti-patterns are all new
features that we propose. Although they are simple, they work well in MVI.
The rest of the paper is structured as follows. Section 2 presents related works. Section 3
describes in detail the specification of the main verb and how the MVI is handled by our
approach. Section 4 gives experimental results. Section 5 presents error analysis and
discussion. Section 6 presents an application of main verb identification, the Chinese Sentence
Breaker. Finally, we draw conclusions and suggest future work in section 7.
2. Related Works
The Chinese language is a typical SVO sequence language, in which ‘V’ is the main verb in
the sentence [Jin and Bai 2003]. The problem of main verb identification has been studied
58 Bing-Gong Ding et al.
extensively by Chinese linguists for a long while [Lv 1980; Ding et al. 1961; Zhang 1982;
Huang 1987; Fan 1995; Liu et al. 2002]. Since the definition of the main verb is related to
different verb-predicate sentence patterns, linguists usually describe the different kinds of
main verbs in the context of verb-predicate sentence patterns.
[Liu et al. 2002] divided verb-predicate sentences into five types: predicates that (1) take
no object, (2) take a single object, (3) take double objects, (4) include an adverbial modifier,
or (5) include complements. Based on this classification scheme, seven specific verb-predicate
sentence patterns were also proposed and discussed individually, including “是” (shi4, to be)
sentences, “有” (you3, have) sentences, series-verb sentences, pivotal sentences, existential
sentences, “把” (ba3) sentences, and “被” (bei4, passive voice) sentences5.
[Fan 1995] introduced a verb-predicate sentence pattern framework that includes seven
subdivided sentence patterns, which overlap with Liu’s classification. For example, SV
(Subject-Verb), SVO (Subject-Verb-Object), SZV (Subject-Adverbial -Verb), and SVB
(Subject-Verb-Complement) patterns in Fan’s framework are similar to (1), (2), (4), and (5) in
Liu’s work. Other sentence patterns include SVL (Subject-Coordination), SCT
(Subject-Series-Verb), and SVD (Subject-Duplicate-Verb). Detailed information can be found
in his book. The reader should be aware that an SVL like “他一边走一边说” (He talks while
walking) or “我们爱祖国爱人民” (We love both our motherland and our people) is equivalent
to a series-verb instead of a sentence with verb-coordination in our definition (see section 3.1
for details). Fan’s and Liu’s works differ in that Fan tries to incorporate more sentence
patterns into a single framework. For example, Fan further subdivides SZV into eight specific
verb-predicate sentence patterns, like “被” (bei4, passive voice) sentences, “使” (shi3, let)
sentences, “从” (cong2, from) sentences, etc. Fan also further subdivides SVB into seven
constructions, like the “verb-resultative construction,” “verb-得construction,” etc.
A particular feature of Huang’s work [1987] is the examples he provides from real texts.
The sentence patterns listed in his work are similar to those in [Liu et al. 2002].
[Zhang 1982] divided verb-predicate sentences into eleven types: verb sentences,
verb-object phrase sentences, verb-compliment phrase sentences, modifier-verb phrase
sentences, series-verb phrase sentences, pivot-verb phrase sentences, series-verb combined
with pivot-verb phrase sentences, “把” sentences, “被” sentences, the negative form of
verb-predicate sentences and the interrogative form of verb-predicate sentences. Similar to
[Liu et al. 2002] and [Fan 1995], Zhang regards the adverbial-modifier as the basis for
subdividing the verb-predicate sentence pattern. However, the author in [Lv 1980] did not use
this kind of basis for classification. In addition, unlike [Lv 1980] and [Liu et al. 2002], Zhang

5 “是”(shi4, to be), “有”(you3, have), “把”(ba3, ba), and “被”(bei4, passive voice) sentences are
Chinese sentences which contain the above words.
Chinese Main Verb Identification: From Specification to Realization 59
uses the whole verb-object phrase, the verb-complement phrase, and the modifier-verb phrase
to subdivide the verb-predicate. However, in our specification, longer phrases or whole
phrases, such as whole verb-object phrases, are recursively defined. This categorization
scheme cannot be used to subdivide our verb-predicate sentence pattern since a shallow parser
cannot provide such information.
The findings in [Lv 1980] were the earliest and most widely ones accepted by other
linguists. According to the different sentence structures, the author in [Lv 1980] introduced 13
types of verb-predicate sentence patterns. See Table 1.
Table 1. Verb-predicate sentence patterns in [Lv]
1. Transitive Verb Sentence
2. Intransitive Verb Sentence
3. Double Object Sentence
4. A sentence whose object is a verb
5. A sentence whose object is a clause
6. A sentence whose object is number
7. A sentence whose object is placed before the predicate
8. “把 (ba3)” Sentence
9. Passive Voice Sentence
10. Complement Sentence
11. Existential Sentence
12. Series Verb Sentence
13. Pivotal Sentence
Transitive Verb Sentence
Subject Adverbial Verb Accusative Non Accusative Auxiliary
modifier Object Ojbect
A 你 从前 学过 英语 吗？
她 唱过 女高音
你 会写 这种笔 吗？
她 最近 吃 食堂 了
B 通县 属于 北京市
这 已经 成为 制度
晚上 不如 早晨
这位同志 姓 李
Figure 1. One example of a verb-predicate sentence pattern

For each type of the sentence, for example the Transitive Verb Sentence shown in Figure 1,
the author of [Lv 1980] provides the predicate in the sentence pattern.
60 Bing-Gong Ding et al.
From the above discussion, we can conclude that when linguists describe and further
subdivide verb-predicate sentences, an important basis of their work is the object of the
predicate. For example, among the thirteen kinds of verb predicates in [Lv 1980], the first
eight kinds of sentence patterns are subdivided according to the type of object. However, our
work is different from theirs because we pay closer attention to main verb types in
verb-predicate sentences than to object types. The reason for this is shown in the following
example.

[MVP 通知/v(tong1zhi1, inform)][NP 他们/r(ta1men2, them)][VP 准备
/v(zhun3bei4, prepare for)][MP三/m(san1, three) 天/n(tian1, days)] 的/u(de)
[NP 干粮/n(gan1liang2, solid food)] 。/ww

See the above example cited from [Meng et al. 2003]. In this example, the sentence is
explained as being a pivotal sentence like a) in [Meng et al. 2003]. Obviously the above
sentence takes more than one parse, such as b), c), and d), if syntactic information only is
available.
a) Pivotal sentence: [Piv-O他们] [Piv-V 准备] 三 天 的 干粮
Note: “他们” is the pivotal object, which acts as both the object of “通知” and the
subject of “准备”.
b) Series verb: [Object 他们] [2nd-V 准备] 三 天 的 干粮
Note: “通知” and “准备” are two series verbs. “他们” acts as the object of “通知”.
c) Clause as object: [Object 他们 准备 三 天 的 干粮]
Note: The whole clause “他们准备三天的干粮” acts as the object of “通知”.
d) Double objects: [Obj1 他们] [Obj2 准备 三 天 的 干粮]
Note: “通知” takes double objects including “他们” and “准备三天的干粮”.
Since it is hard to employ consistent annotation in such sentences and we prefer that our
annotation be theoretically neutral, in our specification, we subdivide a verb-predicate
sentence into four types, including simple verb-predicate sentences, series-verb sentences,
pivotal sentences, and sentences with verb-coordination, instead of using the objects of their
predicates.
The Chinese Penn TreeBank (CTB) is a large-scale bracketed corpus of hand-parsed
sentences in Chinese [Xia et al. 2000; Xue and Xia 2000]. The annotation of the Chinese Penn
Treebank is more complete because they annotate everything, whereas currently we only
annotate verb predicates. Compared with the “Guideline for Bracketing in the Chinese Penn
TreeBank” [Xue and Xia 2000], our specification is different in that the goal of the CTB is to
Chinese Main Verb Identification: From Specification to Realization 61
annotate linguistically-standard and non-controversial parse trees, while the goal of our MVP
annotation is based on chunking which is relatively easily parseable. For this reason, the
guideline of CTB is not entirely identical to our specification. Other differences are listed as
follows.
(cid:122) The annotation of CTB is based on sentences that end with periods, exclamation
marks, or question marks. Our specification defines the main verbs of Chinese simple
sentences (see section 3.2.1 for the reference).
(cid:122) Since we only focus on the output of chunking instead of whole parsed trees as in
CTB, the MVP in our specification is a verb chunk with the main verb, while the
predicate in CTB may be a whole phrase. For example, in CTB, we have the
following:

(IP (NP-PN-SBJ (NR 张三zhang1san1, Zhangsan))
(VP (VV 应该ying1gai1, should)
(VP (VV 参加can1jia1, join)
(NP-OBJ (NN 会议hui4yi4, meeting)))))

“In the above example, the lowest level VP (VP 参加会议) is the predicate,” whereas
based on our parsed chunk results, [VP 应该/v 参加/v] is annotated as an MVP in
this sentence according to our specification.
(cid:122) In CTB, “…a VP is always a predicate, -PRD is assumed…...” However, in our
specification, we only tag the main verb, that is, the verb corresponding to the main
predicate-verb in the sentence. This annotation scheme is consistent to the sentence
analysis methodology of Chinese linguists [Lv 1980].
(cid:122) CTB also tags non-verbal predicates, such as ADJP/NP etc. In our specification, we
don’t consider this case since our focus is verb-predicate sentences.
Linguists provide a grammatical view of Chinese sentences by analyzing them.
Identifying the main verb automatically is a task faced by many computational linguists. Most
of their works have focused on the identification process instead of on the definition of the
main verb. Previous works on MVI can be grouped into three categories: heuristic methods
[Luo 1995; Sui and Yu 1998a]; statistical methods [Chen and Shi 1997; Sui and Yu 1998b];
first heuristic and then statistical methods [Gong et al. 2003].
Heuristic methods were introduced in the early stage of MVI research. Some proposed
approaches depend on linguists’ knowledge; for example, Luo [1995] used hand-crafted rules
to identify predicates. The rules are related to auxiliary words, such as “的(de)” or “得(de)”,
62 Bing-Gong Ding et al.
or to numerical or temporal words. Other approaches employ a bilingual corpus to extract
rules, for example, Sui and Yu’s [1998a] method. However a bilingual corpus is not always
available.
Statistic methods were proposed in [Chen and Shi 1997] and [Sui and Yu 1998b]. Both of
these works are based on verb sub-categorization information. But their categorization
frameworks are different. Chen and Shi’s work [1997] uses only part-of-speech information to
decide on the main verb. Sui and Yu [1998b] use not only sub-categorized part-of-speech
information but also lexicalized context information, such as “的”. Both static and the context
features are integrated into a decision tree model.
[Gong et al. 2003] first used rules to filter quasi-predicates. The features used include the
part-of-speech of the quasi-predicate, the contextual part-of-speech, and the contextual words
like “的”. Then each feature’s weight is calculated from training data. The combined weights
are used to determine the predicates in the sentences.
The works noted above except that in [Chen and Shi 1997] presume that the sentence
boundary has been given. All of them detect predicates in simple sentences. However, they
have a deficiency in that in real text, the sentence boundaries are not provided naturally.
Another difference is that the above works identify verb predicates, nominal predicates and
adjective predicates. In our work, we focus on verb-predicate since both previous [Lv et al.
1999] and our own observations show that the sentences with verb-predicates make up the
most part in corpus.
Another point is that some of the above works use correct verb sub-categorization
information as input [Chen and Shi 1997; Sui and Yu 1998b]. They do not provide main verb
identification evaluation results, where verb sub-categorization needs to be done automatically
as a preprocessing step performed on raw text. Although the task of verb sub-categorization
has long been studied in the Chinese community, the performance achieved has not been
satisfactory. Thus in our work, we make use of more reliable knowledge; for example, we will
provide a closed set of specific verbs whose objects can include multiple clauses, rather than
sub-categorization information in general.
Finally, it is difficult to compare our results with the results of related works because the
test corpora used may be quite different and there are also some differences in the definitions
of the main verb. Thus, we hope that our introduction of a clear specification and corpus for
main verb identification will enable future researchers to compare their results with ours.